Voice Conversion Using Bilinear Model Integrated with Joint GMM-based Classification

Bilinear Model (BM) can express both characteristics within a speaker (style) and phonemes across speakers (content) independently in a speech database. It has a successful application in voice conversion (VC) by extrapolation. However, extrapolation suffers an undesired repetition of BM building and a large-scale estimation of parameters. To tackle these problems, we propose to enhance the normal BM-based VC scheme by integrating a joint Gaussian Mixture Model (GMM)-based classification, assuming that the GMM components correspond to the quasi-phoneme content classes. The enhanced scheme not only optimizes the VC algorithm in computation, but also improves the quality of speech compared to the normal BM-based one, as well as traditional GMM-based mapping system in evaluation experiments.